NuMind (YC S22)

Hello! 📢 Newcomer in the NuExtract family 📢 We just released NuExtract 1.5 smol, based on Hugging Face SmolLM v2 1.7B. A small model, but don't get fooled by its size, the thing can extract! https://lnkd.in/e6-P6jTu In more details: We have been repeatedly asked for small versions of NuExtract. We previously released NuExtract 1.5 tiny, based on Qwen 2.5 0.5B, but unfortunately, we could not get it to: - be multilingual - have continuation abilities (extracting while being given previously extracted information) Recently Hugging Face released SmolLM 2 1.7B, which has excellent benchmark performance, so we decided to give it a go. And it turns out to work pretty well :) We managed to give it all the abilities that NuExtract 1.5 has, and it reaches performance similar to the original NuExtract while being less than half its size! Hope you will make good use of this model, and as always don't hesitate to give us feedback here and in our Discord server: https://lnkd.in/ebAD8NJ6. NuExtract 2.0 is cooking and we want to make it as good as we can! -- model creator: Liam Cripwell, and thanks Leandro von Werra, Daniel V., and Loubna Ben Allal for early access to SmolLM 2 :)

Hi Everyone! NuExtract 1.5 is out! 🎉 TLDR; - Multilingual - Infinite context - Still 3.8B - Still open-source - Beats GPT-4o in English 😳 Hugging Face page: https://lnkd.in/eBjN96Sy Model playground: https://lnkd.in/e77zNHZA Blog post: https://lnkd.in/ehukS_Cm Video presentation: https://lnkd.in/eeYqkSze Congrats to Liam Cripwell and Alexandre Constantin who both worked hard on this project! In more depth: We got energized by the reception of the original NuExtract and decided to pursue in this direction. The main feedback we got was to make it multilingual and to support long documents (the original NuExtract was limited to 4k tokens). So that is what we worked on. We created a multilingual dataset with longer examples and trained phi-3.5 mini on it (128k context). The results are pretty nice. We find that, for English text, NuExtract 1.5 closed the gap with GPT-4o on small documents (~1k tokens), and is quite better than GPT-4o on long documents (~10k tokens)! Non-English results are pretty good as well, but we still have a gap to fill to reach GPT-4o levels (which is allegedly 500x bigger than NuExtract 1.5) We then added an interesting "continuation" ability, which allows NuExtract 1.5 to extract information from a text while being given previous information. This trick allows to process arbitrarily long documents, but more importantly, it allows to keep the memory bounded in order to process very long documents on a small GPU. We detail all of this in the blog post: https://lnkd.in/ehukS_Cm Hope it is useful. Let us know what you think. NuExtract 2.0 is coming and we are looking to improve on every front (and - spoiler alert - to add vision and abstraction abilities). Let's go! -- If you want to be closer to this NuExtract project, you can follow us at NuMind (YC S22) and join our Discord channel: https://lnkd.in/e9F-Qb2Q --

⏳ IT'S THE FINAL COUNTDOWN ! ⌛ 1 week to go until AI for Industry by Artefact. Head to @ Palais Brongniart on September 17th for a full day of conferences on how AI is revolutionising Industry 🏭 🤖 🧠 Check out the program here : https://lnkd.in/eV-texHE 👀 We'll be honoured to welcome to the stage some of the most influential CXOs and thought-leaders on industrial AI : - Jean-Loup Loyer - Chief Data & Analytics Officer @Eramet - Mihir Sarkar - Chief AI Officer @ENGIE - Guillaume Dubrule- Chief Digital and Marketing Officer, Member of the Executive Committee @Rexel - MANGEANT Fabien - Chief Data & AI Officer @Air Liquide - Rodolphe Gelin - Expert Deep Learning and Robotics @Renault Group - Philippe Rambach - Chief AI Officer @Schneider Electric - Vincent ChampAIn - SEVP, Chief IT & Digital Officer and member of the Executive Committee @Framatome - Stéphane PUYDARRIEUX - Group AI Officer, Head of Data Science @Orano - Milton Luaces - Global Vice President of Data for Value Chain @Decathlon - Jean-Vianney Chiron - AI Transformation Manager @Michelin - Caroline Connan - Group Chief Data Officer & Digital Transformation @FORVIA - Vincent Zhen WANG - Chief Digital Officer @Legrand - Sébastien Rousset - Chief AI Officer @NAVAL GROUP - Mikael Volut -  Founder @OSE Engineering and @ÆGIR - Pierre Yves Le Morvan - Industrial and manufacturing lead for France @NVIDIA - Wojciech Janusz - EMEA Data Science and AI for Industry Specialist @Dell Technologies - Yonatan Teboul Group Head of Digital Products and Services @MBDA - Alexandre Bounouh, MBA, PhD, HDR - Directeur Général @CEA-List A big thank you to our partners and exhibitors who will also be showcasing some amazing tech : - yxir - Ask for the moon - Celonis - APREX solutions - Namkin - BtoB Customer Experience - Perception manufacturing - UKG - NuMind (YC S22) - Dawex - Data Exchange Technology And an equally big thank you to our ecosystem partners : - UIMM - PFA - Plateforme automobile - GICAT - GICAN - FIM - Fédération des Industries Mécaniques - GIFEN - FIEEC - A3M - Alliance for Minerals, Metals ores and Metals Message me for the last few passes still available ! 🎟 Shoutout to the Artefact team 🙏 #aiforindustry #industry4.0 #ai #industry

Launching on Product Hunt today, if you could have a look & upvote that would be super useful :) https://lnkd.in/ezcQJqCj

🎉 🎉🎉 NuMind (YC S22) is OUT! 🎉🎉🎉 After a long R&D phase, we are finally coming out of private beta! 😀 You can watch the presentation video here: https://lnkd.in/enpYVaRg And read the release blog post here: https://lnkd.in/eWhdk6nD In a nutshell, 𝗡𝘂𝗠𝗶𝗻𝗱 𝗶𝘀 𝗮 𝘁𝗼𝗼𝗹 𝘁𝗼 𝗰𝗿𝗲𝗮𝘁𝗲 𝗰𝘂𝘀𝘁𝗼𝗺 𝗡𝗟𝗣 𝗺𝗼𝗱𝗲𝗹𝘀 𝗳𝗼𝗿 𝗶𝗻𝗳𝗼𝗿𝗺𝗮𝘁𝗶𝗼𝗻 𝗲𝘅𝘁𝗿𝗮𝗰𝘁𝗶𝗼𝗻. You can use it to create classifiers, NER models (and soon structured extraction models) that are highly tailored to your needs. This means models that are 𝗯𝗲𝘁𝘁𝗲𝗿 and 𝗰𝗵𝗲𝗮𝗽𝗲𝗿 than generic models, and which 𝗰𝗮𝗻 𝗯𝗲 𝘂𝘀𝗲𝗱 𝗶𝗻 𝗮 𝗽𝗿𝗶𝘃𝗮𝘁𝗲 𝘀𝗲𝘁𝘁𝗶𝗻𝗴. The way you create such model is by "teaching the AI", which means telling it what to do and iteratively correcting it until it "gets it". Under the hood, NuMind uses our foundation models (https://lnkd.in/drx4UA3n), some automatic machine learning, and an active learning strategy to obtain good performance with a minimal amount of human effort. We hope that this tool will be useful to you! We are working hard to take this baby way further. We firmly believe that this "AI teaching" process is the way to go and will be the way humans create all sorts of advanced AIs. Feel free to try it out: https://numind.ai/signup, to join our Discord server: https://lnkd.in/ei-ahJ4q, and to give us feedback! Onward to solve information extraction!